DOI: 10.3389/fcomp.2021.624659

One-Liner

Multi-feature late fusion of NLP results (by normalizing text and n-gram processing) with OpenSMILE embedding results.

Novelty

NLP transcript normalization (see methods) and OpenSMILE; otherwise similar to Martinc 2021. Same gist but different data-prep.

Notable Methods

• N-gram processed the input features
• Used WordNet to replace words with roots

Key Figs

New Concepts

• OpenSMILE

A friend recently asked for recommendations for shoes and jackets, and I realized that the links on my gear page has slowly died (very sad). So I figured I should update it with more information and alternatives.

What I (normatively) do

I’ll give specific recommendations shortly, but before I do that I feel like it’d be helpful to give some normative statements about what “good” gear to me feels like.

light, and waterproof, in that order

I try to get things that are both waterproof and light, and if both doesn’t exist (in particular for shoes), I prioritize being light.

Short selling involves betting against the stock.

Process of Short Selling

1. the trader borrows a number of shares from a third party
2. the trader sells them immediately for cash
3. when the security dips, the debt is repaid by repurchasing the same amount of shares of the borrowed security at the lower price
4. traders nets the profit from the negative price differential

If the person shorting

short squeeze

“what happened to GameStock”

sigmoid function is used to squash your data between \(0\) and \(1\). Sigmoid is symmetric. It could take any number and squash it to look like a probability between 0 and 1.

\begin{equation} \sigma(z) = \frac{1}{1+ e^{-z}} \end{equation}

Say you have one discrete variable \(X\), and one continuous variable \(Y\), and you desire to express \(p(x|y)\).

The simplest way to do this, of course, is to say something like:

\begin{equation} P(x^{j} \mid y) = \begin{cases} P(x^{j} \mid y) = 0, y < \theta \\ P(x^{j} \mid y) = 1, y > \theta \end{cases} \end{equation}

Some Ideas

• Error Correction Codes
• Sampling + Quantization
• Compression Algorithms
• Frequency Domain Technologies

Two Main Goals

1. Unit 1: Efficient Representation of Signal (i.e. compression)—we ideally want the smallest sequence of bits to encode the raw signal
2. Unit 2: Preserving Information of Signal (i.e. communication)—we ideally want to communicate our bits while not sacrificing information despite all communication channels being noisy

Unit 1 outline

• compress the same exactly information into less space (lossless compression)
  • what is information (probability and entropy)
  • compression and limits of compression (Huffman Coding)
• removing irrelevant/uninteresting information (lossy compression)
  • key idea: “frequency domain can be aggressively compressed”
  • signals, frequency representation, bandwidth (discrete cosine transform)
  • quantization, sampling, reconstruction (encoding analog signal into digital signal)

Unit 2 outline

• communication basics (channels and noise)
• representing bits for physical/analogue communication (modulation—encoding digital signal into analog signal)
• bandwidth, spectrum shaping/sharing (frequency-domain filtering)
• fundamental limits (channel capacity)
• separation of compression and communication (separation principle)
• adding redundancy to communication schemes (error-correcting codes)

Lectures

Unit 1

SU-ENGR76 Unit 1 Index