Fereshte Khani @fereshte_khani Twitter profile

Pinned Tweet

Fereshte Khani

1 year

Our new preprint proposes a framework for collaborative NLP development (CoDev) that enables multiple users to align a model with their beliefs. An alternative to a single user or a central institution dictating how models should behave!

4

12

93

Last Seen Profiles

@a0sc0

@leemnisa

@lit4tms

@jongjiganjang

@jk_seriouslyjk

@kemirajamiaa

@Rakoongrey

@magialoves2

@iamshaka87

@nick2theallen

@Seihaaaaaa

@EASTLA_NEWS

@iynkyung

@gunthermarder

@DavisManor

@inherentcharlie

@hexensio

@gonukito

@ilyaerkan

@SWGMC

@father

@inyuharea

@LivesOfAsuna

@JolietWestAD

@Noibutsu

@justtakozo

@adammars_

@nyugrossman

@StuartFleetwood

@sarahschauer

@iamOzawaKenji

@SkySportDE

@stw_pdg

@zechariah1680

@YusupHadiansya3

@taynement

Fereshte Khani

@fereshte_khani

6 months

Not the best time to announce this but I joined @OpenAI . Yesterday was my third day 😅

201

92

6K

Fereshte Khani

@fereshte_khani

2 years

In our #Neurips2022 paper, we introduce MaskTune. A method that forces a model to explore new features by masking previously discovered features and finetuning the model over the masked data.

10

71

477

Fereshte Khani

@fereshte_khani

2 years

🚨I’m hiring interns for summer 2023! 🚨 If you are interested in reliable, robust, and fair ML, apply and contact me at fkhani @microsoft .com. Please retweet & share!

9

83

364

Fereshte Khani

@fereshte_khani

2 years

🚨I’m hiring interns for summer 2022! 🚨 If you are interested in fair, robust, and reliable ML, apply and contact me at fkhani @microsoft .com. Please retweet & share!

7

116

288

Fereshte Khani

@fereshte_khani

1 year

🚨 Internship Alert 🚨 There are some interesting internship opportunities on LLMs sponsored by our group, please apply!

4

41

199

Fereshte Khani

@fereshte_khani

2 years

I’ve been asked to give a short presentation about biases in the ML cycle and I thought I would share it here too. 1/n

4

27

148

Fereshte Khani

@fereshte_khani

2 years

It’s been a month since I joined the office of applied research at @Microsoft , and I cannot be happier with my decision. Super grateful to work with these amazing people! @bhecht , @jteevan , @ylongqi , @mengtingwan , @shiladsen , @halfak and Bahar Sarrafzadeh

5

1

115

Fereshte Khani

@fereshte_khani

3 years

1-ML models often rely on spurious features such as backgrounds in images or identity terms in the comments, which undermines fairness and robustness goals. Why do ML models rely on such features even when nonspurious features can determine the target perfectly?

2

18

99

Fereshte Khani

@fereshte_khani

3 years

Ever wondered why removing spurious features (e.g., background) can lead to drop in accuracy? Check our recent paper for the answer. Removing Spurious Features can Hurt Accuracy and Affect Groups Disproportionately #FAccT2021 Joint work with @percyliang

Removing Spurious Features can Hurt Accuracy and Affect Groups...

The presence of spurious features interferes with the goal of obtaining robust models that perform well across many groups within the population. A natural remedy is to remove spurious features...

arxiv.org

0

10

64

Fereshte Khani

@fereshte_khani

3 years

Machine learning models always use features that are imperfect and noisy. But the same amount of noise on each feature should affect everyone equally, right? Check my recent blog post to see how the same amount of feature noise leads to discrimination[1/6]

Blue People v. City of Ney

We show that when ML models use noisy features, they're incentivized to devalue a good score when the candidate is from a lower-performing group.

ai.stanford.edu

3

15

48

Fereshte Khani

@fereshte_khani

3 years

Thank you @trustworthy_ml for highlighting my research. Looking forward to more discussion. If you are interested in these areas, please get in touch!

Trustworthy ML Initiative (TrustML)

@trustworthy_ml

3 years

1/ In this week’s TrustML-Highlight, we are glad to feature Fereshte Khani @fereshte_khani 🎉🎊🎊🎊. Fereshte is a 5th year PhD student at @Stanford , supervised by @percyliang . She is also on the job market this year!!

1

10

43

0

2

41

Fereshte Khani

@fereshte_khani

5 months

I'm presenting this work on Thursday 14 Dec 10:45am! Come and say Hi :-) I also recently become interested in Safe/Constrained RL (especially exploration) and reasoning, so happy to chat about these topics!

1

39

Fereshte Khani

@fereshte_khani

10 months

How can you align machine learning models with the perspectives of multiple users? check my new blog post for some answers 😊

Collaborative Development of NLP Models

There’s an old tale about an elephant in the dark, where people who have never seen an elephant venture into a dark room to touch it, in…

medium.com

1

6

34

Fereshte Khani

@fereshte_khani

3 years

3- In my new blog post, I explain why ML models rely on spurious features even when core features perfectly determine the target; and why accuracy drops after removing spurious features in the absence of the mentioned two reasons.

Removing Spurious Features can Hurt Accuracy and Affect Groups Disproportionately

ai.stanford.edu

1

8

33

Fereshte Khani

@fereshte_khani

1 year

1-The second part of my blog posts on lessons for ML researchers from the legal precedents of anti-discrimination is now live!

1

7

31

Fereshte Khani

@fereshte_khani

8 months

Collaborative development of NLP models got accepted in #NeurIPS2023 🎉🎉 See you in New Orleans!

Fereshte Khani

@fereshte_khani

1 year

Our new preprint proposes a framework for collaborative NLP development (CoDev) that enables multiple users to align a model with their beliefs. An alternative to a single user or a central institution dictating how models should behave!

4

12

93

3

2

28

Fereshte Khani

@fereshte_khani

1 year

1- There are ~90k discrimination charges each year with an average payout of ~300 million. How does the legal process work, and what can machine learning researchers learn from it? check my recent blog post:

1

6

26

Fereshte Khani

@fereshte_khani

8 months

I misread the #NeurIPS2023 rebuttal discussion deadline and couldn’t submit my response to a reviewer who gave us a weak accept because they thought our paper would have a moderate impact. Now that our paper got accepted I put my response here 🙃 In evaluating the impact of our

0

1

25

Fereshte Khani

@fereshte_khani

6 months

❤️😅

OpenAI

@OpenAI

6 months

We have reached an agreement in principle for Sam Altman to return to OpenAI as CEO with a new initial board of Bret Taylor (Chair), Larry Summers, and Adam D'Angelo. We are collaborating to figure out the details. Thank you so much for your patience through this.

6K

13K

67K

1

20

Fereshte Khani

@fereshte_khani

7 months

“Declaration of high confidence mainly tell you that an individual has constructed a coherent story in his mind, not necessarily that the story is true.” -thinking fast and slow The parallel between humans and ML models are very interesting!

2

1

18

Fereshte Khani

@fereshte_khani

1 year

Happy to know that chatGPT knows about taarof 😄

1

2

15

Fereshte Khani

@fereshte_khani

1 year

Stop by @TMLunLimited in #ICLR2023 at 1:40 for my talk on the collaborative development of NLP models Data disparities result in an inaccurate model, similar to understanding an elephant by solely feeling its body.

1

2

14

Fereshte Khani

@fereshte_khani

3 years

4-Note that understanding the reasons behind the accuracy drop is crucial for mitigating such drops. Focusing on the wrong mitigation method fails to address the accuracy drop.

0

13

Fereshte Khani

@fereshte_khani

1 year

I will be in Kigali #ICLR2023 🤩 DM me if you want to meet!

0

12

Fereshte Khani

@fereshte_khani

5 months

@_jasonwei People've been working on selective classification for a long time (). I remember after publishing my first paper on this topic (), I got many emails that why I haven't cited selective classification 😅😅

0

10

Fereshte Khani

@fereshte_khani

3 years

2-One remedy is to remove spurious features to achieve more fair/robust and hopefully more accurate models. But in practice, removing such features reduces the accuracy?!? Maybe the nonspurious features are inadequate, or removing spurious features corrupts the nonspurious ones.

1

0

12

Fereshte Khani

@fereshte_khani

2 years

In the future, GANs might generate breathtaking images, LMs might generate creative stories, ML might even be able to generate stunning music, but we should never forget that ML is NOTHING without data, so the first step should be figuring out how to give credit to data creators.

Sam Altman

@sama

2 years

I think it should also push towards more equality. In the future, I imagine we will need to figure out how to fairly distribute 1) wealth, 2) access to AGI systems, 3) governance decisions about how AGI systems are used.

33

19

208

0

11

Fereshte Khani

@fereshte_khani

1 year

1) What are my goals? 2) Does the current system accurately measure progress toward my goals? 3) Is my strategy effective according to the system's metrics? When I was young I only need to optimize 3 now I need to optimize 1,2,3 together and it is hard!

0

1

10

Fereshte Khani

@fereshte_khani

1 year

respect!

TIME

@TIME

1 year

The Women of Iran are TIME's 2022 Heroes of the Year #TIMEPOY

2K

28K

97K

0

10

Fereshte Khani

@fereshte_khani

2 years

Potential subjects include learning with heterogeneous data (Microsoft has a lot of tenants with different data distributions), making LLMs reliable, and using LLMs to make other models reliable!

1

9

Fereshte Khani

@fereshte_khani

2 years

For example: How can we ensure that the culture of big firms doesn't overshadow the culture of small ones? In contrast to common metrics such as engagement, how can we maximize productivity and organizational success? And many more interesting research questions.

0

9

Fereshte Khani

@fereshte_khani

2 years

My presentation on biases in the ML cycle for the Microsoft Hackathon!

Discrimination in Machine Learning Cycle

Throughout history, people have been discriminated against based on certain characteristics: women were not allowed to vote because of their sex, Black peopl...

www.youtube.com

Fereshte Khani

@fereshte_khani

2 years

I’ve been asked to give a short presentation about biases in the ML cycle and I thought I would share it here too. 1/n

4

27

148

0

8

Fereshte Khani

@fereshte_khani

1 year

Data war? The free usage of individuals' data has now expanded to the free usage of companies' data. With @elonmusk blocking @OpenAI accessing Twitter data, the future of data dividends should be interesting. Great blog post and kudos to ChatGPT for its great answer!

Nick Vincent

@nickmvincent

1 year

After (much) more poking around, I remain very blown away by ChatGPT. But, the idea that people should swap out search engines for language models could seriously erode the sustainability of these systems. Post here: It seems ChatGPT agrees!

2

5

20

1

0

8

Fereshte Khani

@fereshte_khani

8 months

My taste in writing is much higher than my ability to write 🤦‍♀️ #WritingStruggles

0

7

Fereshte Khani

@fereshte_khani

1 year

"Some people say give the customers what they want, but that's not my approach. Our job is to figure out what they're going to want before they do." -Jobs

0

7

Fereshte Khani

@fereshte_khani

3 years

“Simple justice requires that public funds, to which all taxpayers of all races [colors, and national origins] contribute, not be spent in any fashion which encourages, entrenches, subsidizes or results in racial [color or national origin] discrimination.” John F. Kennedy 1963

0

7

Fereshte Khani

@fereshte_khani

2 years

When Marco first shared his template with me, I thought I was wasting my time filling it out, but now I use it for every single project that I do. Great blog posts, highly recommended!

Marco Tulio Ribeiro

@marcotcr

2 years

I never tweet, but here is a blog post I wrote for an intern, may be useful for others too... Part 1: Part 2:

9

128

575

0

7

Fereshte Khani

@fereshte_khani

2 years

The majority of Microsoft customers are other firms so for any ML model we care about each and every user. There are plenty of opportunities for fair, robust and reliable machine learning.

1

0

7

Fereshte Khani

@fereshte_khani

1 year

These internships are part of an amazing program which provides co-mentorship with both a product team and a research team, offering valuable experience and skill-building opportunities in both areas.

0

6

Fereshte Khani

@fereshte_khani

2 years

I'm at #NAACL2022 this week and #ICML2022 next week 🤩 Ping me if you want to talk!

0

6

Fereshte Khani

@fereshte_khani

1 year

We are seeking candidates with expertise in language model augmentation (e.g. retrieval), prompt engineering, hallucination mitigation, techniques for improving inference efficiency, LLM interaction paradigms, and related applied research subjects.

1

0

6

Fereshte Khani

@fereshte_khani

1 year

Our second insight (inspired by Taylor expansion and interpretability works such as LIME) is that we can train a local model specific to the user concept to guide our exploration.

1

0

4

Fereshte Khani

@fereshte_khani

1 year

Do we have an embedding-to-sentence model?

1

0

6

Fereshte Khani

@fereshte_khani

1 year

If you are still searching for an internship and have experience working with LLMs, especially on alignment or interacting with users to elicit their preferences, this is an amazing opportunity for you!

⿻(((E. Glen Weyl/衛谷倫))) 🇺🇸/🇩🇪/🇹🇼 🖖

@glenweyl

1 year

If you are 1) in you're in/around/interested in the Plurality community, 2) have a technical background in GFMs/LLMs especially related to alignment, 3) are available for a summer internship and 4) are in a degree program (ideally PhD), please reach out to glenweyl @microsoft .com

1

14

28

0

5

Fereshte Khani

@fereshte_khani

2 years

Advancement in ML results in models and algorithms that can find “simple functions” that fit training data and generalize very well (likely due to the simplicity of such functions).

1

0

4

Fereshte Khani

@fereshte_khani

3 years

In addition to overcoming the associated effects of being in a group with lower average skills, blue people also need to pass a higher bar to get hired. The more feature noise, the greater the discrimination! [5/6]

1

0

3

Fereshte Khani

@fereshte_khani

2 years

I'm working on biases in data and how they interact with biases in the training. I'm also interested in testing biases and the feedback loops. Please reach out if you want to chat about these topics. n/n

0

5

Fereshte Khani

@fereshte_khani

3 years

Check out our paper @icml2020 for a thorough analysis of how feature noise leads to discrimination and a trade-off between different fairness notions. Joint work with @percyliang [6/6]

Feature Noise Induces Loss Discrepancy Across Groups

The performance of standard learning procedures has been observed to differ widely across groups. Recent studies usually attribute this loss discrepancy to an information deficiency for one group...

arxiv.org

0

1

3

Fereshte Khani

@fereshte_khani

1 year

A great tutorial on alignment!

Jacob Steinhardt

@JacobSteinhardt

1 year

My tutorial slides on Aligning ML Systems are now online, in HTML format, with clickable references! [NB some minor formatting errors were introduced when converting to HTML]

1

7

47

0

4

Fereshte Khani

@fereshte_khani

1 year

Luckily we have good proxies for all concepts (local models) so whenever we make a change we query proxies to see if we create new interference!

1

4

Fereshte Khani

@fereshte_khani

2 years

In our work, in order to avoid shortcut learning we force a model to explore more by blocking the previously discovered patterns. MaskTune results in very good worst-group accuracy without having information about minority groups in training or validation time

1

4

Fereshte Khani

@fereshte_khani

1 year

Joint work with @marcotcr ! Much appreciation to @scottlundberg for all his help and discussion throughout this work. Also thanks to @bhecht and @ZexueHe for their invaluable early feedback. Please reach out if you have any questions/feedback!

0

4

Fereshte Khani

@fereshte_khani

1 year

12/12 - Joint work with @percyliang ! and thank you @AlexTamkin @jmschreiber91 @NeelGuha @PeterHndrsn @megha_byte and @mzhangio for their great feedback!

0

4

Fereshte Khani

@fereshte_khani

2 years

#Mahsa_Amini

Alex Shams

@alexshams_

2 years

In recent days, protests have broken out across Iran condemning the death of a young woman, Mahsa Amini, at the hands of morality police. United in anger at police brutality and restrictive moral codes, Iranians are demanding freedom and an end to government repression:

33

1K

4K

0

4

Fereshte Khani

@fereshte_khani

2 years

💡Fiction idea💡: A world where people of similar age exchange their bodies with each other every 4 years. Do people help each other to take care of their bodies, or do people kill the ones who do not care about their bodies?

0

4

Fereshte Khani

@fereshte_khani

2 years

This study opens interesting future directions: 1) How many “disjoint” explanations can explain data? 2) Besides blocking pixels how else we can define and achieve “disjoint” explanations?

0

4

Fereshte Khani

@fereshte_khani

3 years

In the crop type prediction from satellite images, we got access to new auxiliary data (climate information). How should we best leverage this auxiliary data for better in- and out-of-distribution performance?

Sang Michael Xie

@sangmichaelxie

3 years

🍔🍟"In-N-Out: Pre-Training and Self-Training using Auxiliary Information for Out-of-Distribution Robustness" Real-world tasks (crop yield prediction from satellites) are often label-scarce. Only some countries have labels - how do we generalize globally?

1

37

165

0

3

Fereshte Khani

@fereshte_khani

1 year

A great tutorial on Parameter-efficient Fine-tuning!

Sebastian Ruder

@seb_ruder

1 year

Join our tutorial on Modular and Parameter-efficient Fine-tuning this morning at #EMNLP 2022 w/ @PfeiffJo @licwu . 👨‍🏫 Slides: 📍Location: Capital Suite 7

1

15

107

0

3

Fereshte Khani

@fereshte_khani

2 years

Slides for this presentation: I mostly focus on data collection, training, and feedback loop.

Fereshte Khani

@fereshte_khani

2 years

I’ve been asked to give a short presentation about biases in the ML cycle and I thought I would share it here too. 1/n

4

27

148

0

1

3

Fereshte Khani

@fereshte_khani

1 year

Given these two insights, we prompt LLM such that it generates sentences that proxy and model disagree. Now whenever a user resolves a disagreement we either fix the local model (thus improving user proxy concept) or fix some bugs in the global model.

1

3

Fereshte Khani

@fereshte_khani

2 years

Distribution bias: Historical discrimination creates a gap between the distribution of different groups (e.g., There is a gap between the number of men and women in the Senate). Sampling bias: data can only be sampled from some groups, or data can misrepresent some groups. 6/n

1

0

3

Fereshte Khani

@fereshte_khani

2 years

However, this “simplicity” can be harmful to some groups or force the model to learn shortcuts when there is not enough data to be a paradox for these shortcuts. E.g., see and

2

0

2

Fereshte Khani

@fereshte_khani

2 years

Label bias: for example, historically well-qualified women did not get hired. Feature bias: for example, the number of previous arrests is biased against minorities as it is more probable for them to get arrested with the same amount of illegal drugs.5/n

1

0

3

Fereshte Khani

@fereshte_khani

2 years

Algorithm/Training: Majority bias: ML models usually work better for the majority (e.g., generalization bounds). Simplicity bias: ML algorithms tend to find the simplest model which can cause discrimination for a population with a more complicated function. 7/n

1

0

3

Fereshte Khani

@fereshte_khani

1 year

lol

Ferenc Huszár

@fhuszar

1 year

@russelljkaplan jawdrop

1

8

59

0

1

2

Fereshte Khani

@fereshte_khani

2 years

Business need: Developers' biases can lead to ML models that only address a particular group's need and ignore other groups, or similarly only cause harm to a specific protected group. 2/n

1

0

3

Fereshte Khani

@fereshte_khani

2 years

Data: ML models learn patterns from previously collected data. The data usually reflect the longstanding discrimination against protected groups. This discrimination can be manifested as: 4/n

1

0

3

Fereshte Khani

@fereshte_khani

1 year

2- In the last year of my Ph.D., tired of n different definitions of fairness and m >> n papers on how to optimize them, I decided to read the legal procedure. This blog post is what I wish I had when I began thinking about this issue.

1

3

Fereshte Khani

@fereshte_khani

1 year

CoDev aids users in clarifying their concepts (an area humans often struggle with) and assists ML models to handle conflicts between concepts (an area ML often struggle with due to its inability to accommodate local updates).

1

0

2

Fereshte Khani

@fereshte_khani

1 year

I show how to enable many experts/users to engage with the model and align it with their beliefs while not disrupting one another.

0

2

Fereshte Khani

@fereshte_khani

1 year

But the concept is so big! It might take ages for our random walk to discover areas that aren't aligned with the user (i.e., finding bugs)!

1

0

2

Fereshte Khani

@fereshte_khani

1 year

There are times when I watch a movie or read a book and wish I had done so when I was a child, like when Dorothy tapped her heels in The Wizard of Oz or when I read "All right then, I'll go to hell" in Huck Finn. A Beautiful Day in the Neighborhood was another of those moments.

1

0

2

Fereshte Khani

@fereshte_khani

1 year

6- I expand each step and explain the ML efforts so far and what is missing.

1

0

2

Fereshte Khani

@fereshte_khani

2 years

“You think the writing is communicating your ideas to your readers, it is not! Professional writing is changing their ideas. Nobody cares what ideas you have.”

1

0

2

Fereshte Khani

@fereshte_khani

2 years

Inductive/implicit bias: It is not clear how the unknowns about ML models affect different groups. Misspecification: Not having the true function in the family can affect groups differently. Bias amplification: ML can amplify the biases that exist in the data. 8/n

1

0

2

Fereshte Khani

@fereshte_khani

1 year

4- Despite our desire to have an easy definition for discrimination, proving discrimination is a lengthy process that involves both parties presenting evidence and a judge's decision. This process can provide guidance on how to handle fairness in machine learning.

1

0

2

Fereshte Khani

@fereshte_khani

1 year

It's tricky for users to operationalize their concepts! For instance in the "religion should be neutral" concept, the user might check some sentences but miss others, leading to the assumption that the model is aligned with the concept when in fact it isn't.

1

0

2

Fereshte Khani

@fereshte_khani

8 months

Also, I wrote a blog post about it:

Collaborative Development of NLP Models

There’s an old tale about an elephant in the dark, where people who have never seen an elephant venture into a dark room to touch it, in…

medium.com

0

2

Fereshte Khani

@fereshte_khani

2 years

@DanielKhashabi @jhuclsp @JHUCompSci @JohnsHopkins congrats 🎉

0

2

Fereshte Khani

@fereshte_khani

3 years

The optimal decision for the government (and what the ML models would do) is to choose a higher threshold for hiring blue people compared to red people. In other words, A blue person has a lower chance of getting hired than a red person with the same skill level. [4/6]

1

0

1

Fereshte Khani

@fereshte_khani

2 years

Then he says: And now deer reader, you can do what I have done and it will set you free. But this is not a self-help book!” Stolen focus (), a great book by @johannhari101

0

2

Fereshte Khani

@fereshte_khani

2 years

Problem formulation: The difference between the true need (e.g., content relevancy) and its proxy (e.g., clicked content) might adversely affect some groups. Furthermore, the problem formulation might make it easy for the model to be used unintendedly to harm some groups. 3/n

1

0

2

Fereshte Khani

@fereshte_khani

8 months

@savvyRL I would love to! I’ll ping u after I finish the camera-ready 😊

0

Fereshte Khani

@fereshte_khani

1 year

Our first insight is that we can use LLM to explore a concept by repeatedly using a set of 3-5 sentences from that concept as a prompt and asking for similar examples to be generated. This technique enables us to perform a random walk within the user's concept.

1

0

2

Fereshte Khani

@fereshte_khani

6 months

"With no effort, he had learned English, French, and Latin. I suspect, however, that he was not very capable of thought. To think is to forget differences, generalize, make abstractions. In the teeming world of Funes, there were only details, almost immediate in their presence."

0

2

Fereshte Khani

@fereshte_khani

2 years

A union for training data 😁

0

2

Fereshte Khani

@fereshte_khani

1 year

Interference is another big hurdle in ML! It's impossible to alter one part without affecting the rest!

1

0

2

Fereshte Khani

@fereshte_khani

2 years

“The point systems often don’t measure what we want to measure. They artificially simplify or distort our values. Our goals are messy, complex and multifaceted, but then they are collapsed down to scoring systems.”

1

0

2

Fereshte Khani

@fereshte_khani

2 years

Implicit change: Model predictions change the incentives for individuals (e.g., if the probability of arrest is high without committing the crime for one group, members of that group become incentivized to commit the crime regardless). 14/n

1

0

2

Fereshte Khani

@fereshte_khani

2 years

Explicit changes: Models can also alter human behavior more explicitly (e.g., a college-admitted individual goes through training which increases her skill level, or a recommendation system can alter a person’s food choices by recommending many types of junk food). 15/n

1

0

2

Fereshte Khani

@fereshte_khani

1 year

7- One of the interesting pieces of evidence for proving disparate treatment is direct evidence of motive which is any statement/act by the employer that indicates a bias against members of a protected group.

1

0

1

Fereshte Khani

@fereshte_khani

2 years

@jteevan @marknb00 @tanzeemc @nkittur @amyjko @MarilynTremaine @roelvertegaal Congrats! 🎉

0

1

Fereshte Khani

@fereshte_khani

2 years

“A writing teacher should not ask a student: why do you think that? He should ask: why should I think that?"

1

0

1

Fereshte Khani

@fereshte_khani

1 year

"I think that it's much more dramatic that two men could be working out their feelings of anger -- much more dramatic than showing something of gunfire"

1

0

1

Fereshte Khani

@fereshte_khani

2 years

“If this was a self-help book, I would be able to serve up a delightfully simple conclusion to this story. Those books have a very satisfying structure. The author identifies a problem—usually one he’s had himself– and he talks you through how he personally solved it.

1

0

1

Fereshte Khani

@fereshte_khani

3 years

The government wants to hire anyone with a skill level above 0. To do so, the government takes an exam from everyone, which is a noisy indicator of people's skill level (the noise is the same for everyone). [3/6]

1

0

Fereshte Khani

@fereshte_khani

2 years

Models change over time as well to fit the population (e.g., A/B testing for model selection). Although any model that incorporates user feedback changes over time, ML makes the process faster by training the models rapidly using the newly collected data. 16/n

1

0

1

Fereshte Khani

@fereshte_khani

2 years

@aidaa @MohammadAtari90 @berendennedy congrats 🎉

1

0

1

Fereshte Khani

@fereshte_khani

3 years

@2plus2make5 Interesting that you all believed her narrative. I thought she is experiencing some sort of AI-Schizophrenia, and none of these have happened in the real world.

1

0

1