🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Spandan Madan on generalization in models of vision. (presented at NeurIPS 2024). youtu.be/pBCMr9pAHaw

Are DNNs good models of the brain? We answer this question in our recent work: openreview.net/forum?id=kD1kp…. Had a lot of fun presenting this at NeurIPS 2024! The associated interview with CBMM can be found here: youtube.com/watch?v=bhXB-d….

We all know about adversarial attacks. But is a malicious adversarial agent even necessary to attack these models? Not quite. In our recent #Neurips2024 work, we show networks are highly susceptible to data hiding in plain sight within the training data distribution!

Why do vision models struggle to generalize? Our #NeurIPS2024 paper identifies the underlying cause---computer vision models simply need to be trained with data similar to biological vision. We present the Human Visual Diet dataset, and models that can leverage this data.

Very proud that our work was covered by CBMM and Harvard SEAS (seas) news! In a nutshell, we explore the perils of Neural Network based models of the brain! cbmm.mit.edu/video/benchmar… seas.harvard.edu/news/2025/01/e…

If you're an undergrad with strong programming skills, and would like to work on LLMs + Human Memory at Harvard University, drop me a message.

Super proud of the amazing work led by serena bono, which just made it to the homepage of MIT: mit.edu. Associated News: news.mit.edu/2025/new-train… Video Interview explaining the project: youtube.com/watch?v=JEyP1v…

Built a new benchmark evaluating OpenAI O3 and Google Gemini 2.5 Pro on cognitive tasks. Google beating OpenAI hands down. Every cognitive task. More to come soon.

Going to be honest. I have a PhD in Machine Learning, and I have never ever used Discrete Structures. Make that leap. Skip those stairs.

Huh. Looks like Plato was right. A new paper shows all language models converge on the same "universal geometry" of meaning. Researchers can translate between ANY model's embeddings without seeing the original text. Implications for philosophy and vector databases alike.

All models are the same. Please enjoy each model equally.

Interesting problem now that LLMs have memory across chats---Claude 4 copied and replaced line of code from an older file version from a past chat. Took forever to track.

The same might be true for code. The skill of thinking "what to code" might get atrophied. Seeing it in younger students already.

Couldn't agree more. In my entire PhD---new architecture = 2-3% gains, better data = 20-40% gains.

Calling c the "speed of light" completely misses the point. Rather, c is the "spacetime exchange rate": how many units of space you can exchange for one unit of time. In actuality, everything travels at the "speed of light", just not necessarily through space alone... (1/4)

claude: “you’re absolutely right!” me:

Learnings from the Bay Area---Internet really sucks at Stanford University.

Hot take---OpenAI will name it GPT-3.14.

When p<0.05, even though sample size < 10.