The Future Isn’t Scary | Self-Driving Car QQQ ad lying car.
I hereby solicit suggestions for the video of the day. Please email me your ideas with explanations. Selected entries will win 5 homework points. If your video is played at the beginning of class, you must also briefly explain something about the video and something about yourself - in person.
Tune in for a presentation from the Yale Office of Career Strategy. This talk will review a number of tips and advice on the internship search process for CS jobs. There will also be a panel of students who have obtained internships at Microsoft, Netflix and Amazon to speak about their experiences and offer advice. Open to all students!
Title: Robust NLP: Can We Do Better Than Bigger?Abstract: The robust text understanding and generation capabilities of today’s NLP has been driven by the assumption that larger models and larger datasets improve predictive performance. However, this approach may not be sufficient to address real-world challenges, particularly in low-resource languages and domains.
In this talk, I will argue that robust NLP requires us to move beyond accuracy and perplexity as evaluation metrics. Towards this goal, I will present the idea of learning from knowledge rather than data, which exploits input and label meaning, in the presence of invariant domain knowledge. This approach generalizes traditional machine learning and has shown promise across multiple NLP problems. I will connect back to the theme of low-resource domains by presenting a text-based crisis counseling application. I will conclude by outlining future research directions around the theme of making inferences about text across domains despite limited data and compute resources.
Bio: Vivek Srikumar is an associate professor in the Kahlert School of Computing at the University of Utah. His research lies in the areas of artificial intelligence, natural language processing and machine learning, and has been primarily driven by questions arising from the need to efficiently reason about textual data with limited supervision. His research has been published at various AI, NLP and ML venues, and has been recognized by a paper award at EMNLP 2014, and honorable mentions from CoNLL 2019 and the IEEE Micro magazine. His work has been supported by research grants from NSF, US-Israel BSF, NIH, and awards from Google, Intel, Nvidia and Verisk. He has served as associate program chair of AAAI 2022 and the program co-chair of CoNLL 2022 and ACL 2024. Furthermore, he has organized several workshops hosted at the primary ML and NLP conferences around the theme of how learning and structured knowledge intersect. He was a post-doctoral scholar at Stanford University before moving to Utah, and prior to that, in 2013, he obtained his PhD from the University of Illinois at Urbana-Champaign.
Website: https://svivek.com
An Interdisciplinary Symposium based on previous work you have done in this area. We aspire to make this a truly interdisciplinary event by offering a variety of formats including panels, lightning talks, posters, art works, performances, and demonstrations (robots welcome).o Event Date and Location: May 9, 2025, Kline Tower
o Visit the Symposium Website
o Deadline: Submit a 150- to 250-word proposal abstract by March 31, 2025.
We observe that Senators and Supreme Court justices often give explanations or opinions that hide or obscure their real reasons. That is, they lie. Should a computer lie? or more realistically, When should a computer lie? We call your attention to HAL from 2001: A Space Odyssey.
Are These 10 Lies Justified? from The New York Times, December 14, 2015. See comments as well.
Using polleverywhere, list (by number) which of the ten lies from the NYT article you believe are justified. https://pollev.com/slade
Results:
all except 1, 2
1, 4, 5, 6, 7, 8, 9, 10
1, 5
1,2,5,6,7,8,9
9,10
1, 5, 7, 9, 10
All but 6, 10
1, 5, 8, 9
1, 2, 9
1, 5, 7, 8, 9, 10
5
1,3,4
1,2,3,4,5,7,8,9,10
1
>>> x = [1,2, 1, 4, 5, 6, 7, 8, 9, 10, 1, 5, 1,2,5,6,7,8,9, 9,10, 1, 5, 7, 9,
10, 1,2,3,4,5,7,8,9, 1, 5, 8, 9, 1, 2, 9, 1, 5, 7, 8, 9, 10, 5, 1,3,4,
1,2,3,4,5,7,8,9,10, 1]
>>> len(x)
61
>>> count = {}
>>> for i in range(10):
... count[i+1] = 0
>>> for i in x:
... count[i] += 1
>>> count
{1: 12, 2: 5, 3: 3, 4: 4, 5: 9, 6: 2, 7: 6, 8: 6, 9: 9, 10: 5}
>>> for x in count:
... print (x, count[x])
1 12 # surprise party
2 5 # plane crash
3 3 # father with dementia
4 4 # husband drowned in car accident
5 9 # real estate testers
6 2 # Cuban missile crisis
7 6 # family planning hiring question
8 6 # placebo surgery
9 9 # used car salesperson
10 5 # praising children
In traditional programming, we write rules that work on data to produce answers. The rules are the program and the data may be images or search terms or the web pages populating the internet. The results are answers: a set of links to relevant web pages.
Traditional programming paradigm: Input: rules + data Output: answers
With machine learning, we alter that paradigm. The input to machine learning, at least for supervised learning, is data and answers. The output is the rules for prediction.
Machine Learning paradigm: Input: data + answers Output: rules (prediction function)
Consider the hw6 dataset of Titanic passengers. The data is the training set of passengers with their various features. The answers are the target values, 0 or 1 to designate perished or survived. We process that data and answers through some learning algorithm, and get a prediction function which is a set or rules.
We are trying to get the machine to figure out the rules, based on the data. That is machine learning.
By and large, we do not make any cognitive claims for machine learning. ML does not provide models for human cognitive processes. Even neural networks, which are inspire by the architecture and behavior of neurons in the brain, should not be construed as cognitive models.
ML is a pure engineering or statistics endeavor. We are interested in the input/output behavior of the machine. We are not trying to understand how people learn.
Having said that, there are cognitive models of human learning. The reality is that they do not scale very well. They work on small domains, but cannot handle boatloads of data that are the wheelhouse of ML.