Blog

“…Do you see what a captious argument you are introducing — that, forsooth, a man cannot inquire either about what he knows or about whit he does not know? For he cannot inquire about what he knows, because he knows it, and in that case is in no need of inquiry; nor again can lie inquire about what he does...

Memory cannot be understood, either, without a mathematical approach. The fundamental given is the ratio between the amount of time in the lived life and the amount of time from that life that is stored in memory. No one has ever tried to calculate this ratio, and in fact there exists no technique for doing so; yet without much risk...

We investigate the neural dynamics in an exicitatory-inhibitory neural nerwork (E-I network) with ReLU activation and Hebbian learning. By local bifurcation analysis, we find there are ten possible neural dynamical patterns between one excitatory neuron and one inhibitory neuron with all four types of synaptic connections. The unstable node never exists in this system, and when there is no external...

In this work, we tackle the problem of real-time virtual reality streaming in low-bandwidth network conditions. In low-bandwidth settings, VR streaming faces an inherent trade-off between responsiveness and image resolution, both of which are crucial to a high quality user experience. To overcome this trade-off, we propose to send low-resolution frames over the network and then apply neural super-resolution on...

Mitochondria play a crucial role in the functioning of neurons by synthesizing ATP and buffering intracellular Ca2+, both necessary for synaptic function. The morphology and location of mitochondria can provide insights into specialized function they perform in different parts of the neuron. We reconstruct 81,397 mitochondria within 180 pyramidal neurons segmented and skeletonized within an electron microscopy volume of mouse...

This research is driven by our curiosity about the limits to which the current technology for natural language processing could be pushed: if humans are capable of reading jumbled sentences, can machines do? To answer this question, we evaluate the ability of state-of-the-art models to recognize and comprehend wordand character-level jumbled sentences, analyze factors that may influence machine’s performance, and...

This research project studies a novel online decision-making problem of choosing the correct action each round given access to experts that are knowledgeable, rational, and truthful, rather than arbitrary. We ask, in this setting, whether we can surpass the best expert using voting-based algorithms when the experts provide additional information along with their votes. We first show that we can...

We propose a novel approach of imitation refinement, which improves the quality of imperfect patterns, by mimicking the characteristics of ideal data. We demonstrate the effectiveness of imitation refinement on two real-world applications: in the first application, we show imitation refinement improves the quality of poorly written digits by imitating typesetting or well-written digits. In the second application, we show...

DeepMind researchers claimed state-of-the-art results in the Arcade Learning Environment in their recent ICML paper “A Distributional Perspective on Reinforcement Learning”. They investigate a distributional perspective of the value function in the reinforcement learning context and further design an algorithm applying Bellman’s equation to approximately learn value distributions, which results in better policy optimization. The discussion in their paper follows...

Nowadays, crowdsourcing becomes an economical means to leverage human wisdom for large-scale data annotation. However, when annotation tasks require specific domain knowledge that people commonly don’t have, which is normal in citizen science projects, crowd workers’ integrity and proficiency problems will significantly impair the quality of crowdsourced data. In this project, I focused on improving the crowd workers’ reliability during...

This article is about the theoretical derivation of Policy Improvement Bound and practical policy optimization algorithms discussed in the paper Trust Region Policy Optimization (Schulman, et al, 2015). TRPO is an interesting idea which optimizes policies with guaranteed monotonic improvement. In theory, its algorithm design looks elegant and justified. In practice, it performs robustly on a wide variety of tasks....