How we search for better deep learning algorithms

Deep learning systems are built from small building blocks stacked into an algorithm. Choosing which blocks to use, how deep they go, and how they train is hard to do by eye.

Instead of freezing one algorithm up front, we employ a meta-search. That is a search that invents and tests many candidate algorithms automatically. Each candidate gets a short training pass. Then a locked score ranks it against the others. We applied that search in DenoiseOpt.

The goal is simple to say and hard to do well: find algorithms that score high without cheating the measure, and without needing a person to tune every trial. The rest of this article walks through the score we locked, the search styles we compared, and what the numbers do and do not show.

DenoiseOpt repairs the click in looping sounds, explained in how to remove the click in a wavetable loop. A normal repair tool is one algorithm with a few knobs. DenoiseOpt works one level up. It searches over the repair programs themselves.

There are two loops. The inner loop takes one candidate repair and fits its weights for a few steps. The paper calls this a FitCell. The outer loop decides which candidate to try next. It can change the depth and width of a repair, swap building blocks, or change training settings. That outer loop is the meta-search.

The building blocks are small. The search can pick from 17 block types, including simple dense layers, small convolutions, attention and LSTM cells. The goal is a tiny repair that runs fast, so the search never considers large audio models.

A search is only as good as its score. If the score has a loophole, the search will find it. R_blend is designed to close the two obvious loopholes.

The score uses a test pair from a sound generator. For each random seed it makes the same cycle twice. One copy has a broken wrap, the cracked cycle. The other copy is identical but without the break, the ideal sibling. The repair only ever sees the cracked copy. The ideal sibling is used for scoring and nothing else. No human labels or studio recordings are involved.

Each cycle is repeated 16 times before scoring, so faults that only appear on repeat get caught. Then the score looks at two zones.

The seam: 8 samples at each head and tail of the cycle. Here the repair should match the ideal sibling. This part is R_seam.

The body: everything else. Here the repair should leave the cracked input as it was. This part is R_body.

Both parts use the same simple measure. It compares the error to the size of the reference signal and maps the result to a value from 0 to 1. Then R_blend = 0.7 × R_seam + 0.3 × R_body.

This closes both loopholes. A repair that does nothing scores badly on the seam. A repair that smooths the whole cycle scores badly on the body. The weights and the formula were locked before the reported experiments, so the method cannot move its own goalposts.

R_blend says how good a repair is. The search also needs to care about speed, because a synth has to run the repair in real time. So the search ranks candidates by J: J = R_blend − 0.02 × latency_norm.

Latency_norm grows with the repair's run time on a log scale. It reaches 1 at 50 milliseconds. The penalty is small on purpose. It only breaks ties between repairs of similar quality.

The paper keeps the roles apart. R_blend is used for all tables and figures. J only picks the champion. A frozen Noise2Noise neural model serves as a pass or fail gate on held-out data.

The outer loop rotates through several branches. Each one proposes new candidates in a different way.

Genetic algorithm (GA): keeps the best candidates, mixes them and mutates them.

PPO: a reinforcement learning method that learns which edits to a repair tend to raise the score.

Population-based training (PBT): copies settings from the top quarter of the population into weak members and nudges them.

Architecture moves: change depth, width and block mix.

When the best score stops improving for a while, the loop shifts its branch mix. The paper says plainly that it reuses these known tools. It does not claim a new search or reinforcement learning algorithm.

This is the least obvious part. On the generator, a cracked cycle with no repair already scores close to 0.97, because only the seam is broken. Good repairs sit in a narrow band just below 1. The differences that matter are a thousandth to a hundredth of a point.

Raw scores that close together give the learner almost no signal. So PPO does not see R_blend directly. It sees a shaped reward, for example the gain over a classic cosine crossfade. PBT is allowed to switch between four reward modes while the search runs. The paper treats reward shaping and search settings as tunable parts of the method, on equal terms with the repair design.

The paper ran six outer-loop families with the same budget of 5,000 evaluations and three search seeds each: random search, CMA-ES, plain REINFORCE, aging evolution, a TPE Bayesian search and the hybrid. Each used the same short FitCell and the same scores.

The hybrid had the highest mean R_blend, 0.9748 ± 0.0018. The paper calls this selection on three seeds. It ran no ANOVA, so the ranking is a guide and not a significance result.

On the five-seed holdout, a short hybrid run reached 0.9708 ± 0.0014. The frozen Noise2Noise model scored 0.9767 ± 0.0015, so it stayed ahead. A two-sided cosine crossfade scored 0.542 ± 0.038.

Meta-search: a search over many repair programs. Tuning one fixed program is the usual approach.

FitCell: the short inner training run that fits one candidate before it is scored.

R_seam, R_body, R_blend: the seam match, the body match, and their 70/30 mix.

J: the ranking score used to pick a champion. R_blend minus a small speed penalty.

Reward shaping: turning a raw score into a signal a learner can use, for example by measuring the gain over a baseline.

No listening tests. The paper reports zero listeners and no MOS or MUSHRA scores. R_blend is the primary evidence.

R_blend is an objective proxy. A high score means the samples match the target. It does not prove the result sounds better.

The target comes from a generator. Real recordings have no ideal sibling, so the score cannot be computed on them the same way.

The search comparison rests on three seeds. The single run that cleared the Noise2Noise gate (about 0.9912 vs 0.9750) is one search seed and is marked provisional by the paper.

The paper is a preprint and has not been peer reviewed.