My Dad works at a crappy job, drinks too much, and his only friend is his brother.

I always thought he led an unfulfilled life.

But I had an epiphany:

Important paper from Google on large batch optimization. They do impressively careful experiments measuring # iterations needed to achieve target validation error at various batch sizes. The main "surprise" is the lack of surprises. [thread]

https://t.co/7QIx5CFdfJ

The paper is a good example of lots of elements of good experimental design. They validate their metric by showing lots of variants give consistent results. They tune hyperparamters separately for each condition, check that optimum isn't at the endpoints, and measure sensitivity.

They have separate experiments where the hold fixed # iterations and # epochs, which (as they explain) measure very different things. They avoid confounds, such as batch norm's artificial dependence between batch size and regularization strength.

When the experiments are done carefully enough, the results are remarkably consistent between different datasets and architectures. Qualitatively, MNIST behaves just like ImageNet.

Importantly, they don't find any evidence for a "sharp/flat optima" effect whereby better optimization leads to worse final results. They have a good discussion of experimental artifacts/confounds in past papers where such effects were reported.

My System for momentum trading🧵:

1. Stock Identification:

Daily and Weekly Delivery Analysis using @mystockedge
Theme Confirmation using Internet and Network

2. Stock Confirmation:

Relative Strength of a stock visavis market using RS55 model listed in @mystockedge

1/n

3. Stock Participation:

Only when chart showing short term breakout along with RS55>0 and RSI>50. I am ready to leave some initial movements.

4. Position buildup:

Every time the stock gives a breakout from the previous swing high add more position and go all-in.

2/n

5. Stock Exit:

I run two portfolios: Short term Swing and Long term Swing.

For Exit in Short term swings, I use a 2hourly Chart with super trend and RS. I am using @tradingview for the same.

For long term swing, I exit when the RS of the stock goes below 0. Period.

n/n

A thread on HN about bad code in legacy projects both makes me think how little we've learned as a discipline over the years and, honestly, how little credit we give ourselves for some pretty major

Fun going down this list and thinking: "Hmm, plausible at a well-run modern software shop", "Hmm, possible, but requires implausible tradeoffs", "Literally disallowed by languages", and "If you were to attempt doing that our test suite wouldn't let you merge."

I think we as an industry celebrate (not quite the right word) failure too much and don't celebrate success nearly enough. There is no DailyWTF for competent execution, word of which generally stays pretty local to the source while incompetence passes into legend.

Alrighty let me try to thread the needle on being the change I want to see in the world while not giving away anything that will get me in trouble:

Ruby has wonderful developer ergonomics. Typed languages are easier for machines to guarantee the correctness of. We built a type checker for Ruby (and I believe it is slated for OSS release sometime).

c.f.