Ahmed Imtiaz Humayun

Hi this is Imtiaz! I am currently a Research Scientist at Google working on Gemini Post-Training. I completed my PhD at Rice University with Dr. Richard Baraniuk, during which I also spent a year and a half as a Student Researcher at Google Research.

I like thinking about how data and objectives shape training, memorization and learning dynamics in neural network based models and agents. Individual neurons in any deep neural network learns to induce changes or non-liniearities in the function being learned. A lot of my research involves studying how the local geometry of such changes or non-linearities relate to memorization, generalization, and robustness in DNNs. My work has broad implications across domains, e.g., in learning theory, interpretability, and safety. Apart from this I also like thinking about ways to undo the changes learned by neurons after training, i.e., unlearning and the effects of looping in the outputs generated by neurons as their input, i.e., synthetic data training.

I am a recipient of the Loewenstern Fellowship at Rice. My work has been published at NeurIPS, ICML, ICLR, CVPR, ICASSP and INTERSPEECH to name a few, and has been featured in multiple news sources, e.g., The New York Times, Communications of the ACM, The Telegraph, yahoo!finance, FORTUNE, WIRED, New Scientist, Futurism, Tom's Hardware, The Daily Star, The Business Standard, Prothom Alo (Bengali), and IEEE Signal Processing Magazine. Outside of research, I have founded Bengali.AI, a non-profit research initiative that crowdsources vision-language datasets and open-sources them through benchmarking competitions, e.g., $50,000 USD Out-of-Distribution Speech Recognition Comp @ Kaggle

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

We show that Grokking, a perplexing phenomenon in deep neural networks (DNNs), manifests for adversarial examples across various practical settings like Resnets on Imagenette and GPT on Shakespeare Text. The emergence of delayed generalization and robustness is explained by a phase change in a DNN's mapping geometry, when a robust partitioning of the input space by the DNN emerges.

We provide quantitative and qualitative evidence showing that for models ranging from toy settings to foundational Text-to-Image models like Stable Diffusion 1.4 and DiT-XL, the local geometry of the generator mapping is indicative of downstream generation aesthetics, diversity, and memorization. Finally we demonstrate that by training a reward model on the local geometry, we can guide generation to increase the diversity and human preference score of samples.

For low dimensional regression tasks such as fitting implicit neural representations on 2D/3D signals, we present a simple pre-training method requiring only 10 regression tasks/signals, to obtain weight initializations that always results in faster convergence upon fine-tuning.

We adapt current scaling laws with the average active parameter count to obtain scaling laws for sparse pre-training. We also present pruning routines for sparse pre-training that achieves the same performance as dense pre-training for a fixed FLOP budget. This results in significant reductions in inference compute.

We present a novel algorithm for self-improving diffusion models using their own generated samples. By, fine-tuning a base model on its own synthetic data, we obtain a collapsed/MAD score function, that we use to negatively guide generation for the base model. This results in mitigation of model collapse and (self-)improvement of generation performance (FID).

We study the phenomenon of training new generative models with synthetic data from previous generative models. Our primary conclusion is that without enough fresh real data in each generation of a self-consuming or autophagous loop, future generative models are doomed to have their quality (precision) or diversity (recall) progressively decrease.

The first provably exact method for computing the geometry of ANY DNN's mapping, including its decision boundary. For a specified region of the input space, SplineCam can be used to compute and visualize the 'linear regions' formed by any DNN with piecewise linear non-linearities, e.g., LeakyReLU, Sawtooth.

Using spline theory, we present a novel method for imposing analytical constraints directly on the decision boundary for provable robustness. Our method can provably ensure robustness for any set of instances, e.g. training samples from a specific class, against adversarial, backdoor or poisoning attack.

A provable method for controllable generation based on quality and diversity from any pre-trained deep generative model. We show that increasing the sampling diversity helps surpass SOTA image generation.

A novel and theoretically motivated latent space sampler for any pre-trained DGN, that produces samples uniformly distributed on the learned output manifold. Applications in fairness and data augmentation.

Using spline theory, we present a method for exact visualization of deep neural networks that allows us to visualize the decision boundary and also sample arbitrarily many inputs that provably lie on the model's decision boundary

Repeated samples and sampling bias may manifest imbalanced clustering via K-methods. We propose the first method to impose a hard radius constraint on K-Means, achieving robustness towards sampling inconsistencies.

We extend existing methods that rely on the use of kernels to variable-length sequences by using the Recurrent Neural Tangent Kernel (RNTK).

We show that novel Convolutional Neural Network (CNN) layers that emulate different classes of Finite Impulse Response (FIR) filters can perform domain invariant heart sound abnormality detection.

Very Deep Convolutional Residual Network achieve state-of-the-art results in sleep staging, using only raw single channel EEG.

Convolutional Recurrent Neural Networks outperform SOTA RNN based compression methods as well as JPEG 2000 for X-ray image compression.

State-of-the-art heart abnormality classification using an ensemble of Representation Learning methods.

We propose novel linear phase and zero phase convolutional neural networks that can be used as learnable filterbank front-ends.

IEEE Signal Processing Cup Honorable Mention for Real-time Music Beat Tracking Embedded System.

Bengali.AI is a non-profit in Bangladesh where we create novel datasets to accelerate Bengali Language Technologies (e.g., OCR, ASR) and open-source them through machine learning competitions (e.g., Grapheme 2020, ASR 2022)

Jointly largest open-sourced Bengali ASR dataset as well as first Bengali Out-of-Distribution Speech Recognition benchmarking dataset.

First Multi-Domain Bengali Document Layout Analysis Dataset, with 700K polygon annotations from image captured documents in the wild.

We have crowdsourced the first public 500 hr Bengali Speech Dataset on the Mozilla Common Voice platform, with speech contributed by over 20K people from Bangladesh and India.

A benchmark datset for multi-target classification of handwritten Bengali Graphemes, with novel implications for all alpha-syllabary languages, e.g., Hindi, Gujrati, and Thai.

The first large scale Multi-Domain Bengali Handwritten Digit Recognition Dataset

Yet another steal of Jon Barron's amazing website.