Mojtaba Soltanalian

The journey begins with a familiar parable: A group of villagers hears that a strange and enormous creature from India has arrived. Eager to understand it, they sneak into the dark barn where it’s kept. One touches the trunk and says, “It’s like a snake!” Another, feeling a leg, declares, “No, it’s like a tree trunk!” A third, touching an ear, insists, “It’s like a fan!” Who is right? The truth is, none of them sees the whole elephant. But what if they had more points of contact? Even if each touch lacked fine detail, the abundance of samples would provide a much better picture of the creature; this is the magic of one-bit sampling. Another core insight is the sample abundance singularity: We have shown in our prior work that, under quantization, increasing data volume can reduce complexity in a counterintuitive manner. Finally, the talk explores inference models that trade bit-depth for bit-width; i.e., decreasing the bit-per-link rate but simultaneously increasing the number of layers for more sophisticated representation (noting that neural networks, in essence, are approximators or “samplers” of an exact mapping).