Ever notice that sometimes the neural networks on this blog do a better job of imitating weird datasets than at other times? Here are two major things that affect how convincing a neural network version will be:

1. Amount of data vs complexity of the problem

2. How easy the dataset is to fake

So, the neural network’s beer names were usually indistinguishable from human-generated names: I had lots of examples, and the names are silly to begin with. The neural network had about the same number of examples of first lines of novels, but these were longer and it had to learn to use grammar correctly as well. Most of its attempts were off in some way.

Here’s one case in which “easy to fake” saves the day: generating thesis titles.

Andromeda Yelton, a software developer at MIT Libraries, sent me a list of 44,126 thesis titles from MIT (you can read them here). Since a thesis represents the culmination of years of deep study of a particular topic, their titles are usually highly specialized. The title of mine, for example, was “Design of Robust, Low-Threshold Wavelength-Scale Nanolasers using Optical and Thermal Analysis”. Each word is chosen with care, but only someone with knowledge of the field would notice a mistake. This makes thesis titles delightfully easy to fake.

Here’s a typical example of neural net-generated thesis titles.

Reconstruction of metal-to-motion : construction of plasma emissions

Computational approaches for distributed microscopy

Optical effects on virtual radio Projects

Experiments, and protein games from multiple atoms

Hydrogel wireless charging via nanoparticle education

Supersone legal questions support kits regulation on qubits and transportation

Atoms and characteristics of monolithic nanocity

Sometimes it gets easier to spot a fake when the neural net tries to mix together two different fields of engineering.

Spacecraft Coal battery induced by mortgage microcontrol

Integrated silicon contact and n-type underwater vehicles

Extreme polymer systems in policy and permanent markets

Privatization algorithms for meyout quaternion visibility motors


A diffusion list over galaxy effects in a new control of Long Turbines

Finite interpretation of nucleir Lncuezzji blocks on compliants and their effects on renform avoidance and environmental funding of a circuit

The neural net flagrantly makes up sciency-sounding words.

An amundium-based solution of current learning on drip power control of pressurized microenvironmental energy construction

Robustness and reality horistic kneehofficiation cost of the cities energy spectrometer

Using Breakozoic : societial submaries within the molybdenum limbters

Rural extension of sralinium finite biomimetic analysis : Dehlum-jet price buildings

Critical emerging paralusion and North Atlantic : rehabilitation of endetential biomimetic nitrogen finage analysis, building teats by vesicle plant nanoscale platinuse-time demiler system dynamic to polypeptions under city, and black vortiis

If it makes up too many words, it may end up confusing even itself.

Modeling of the prosthesis design of space structures for intelditions and connection applications in nanostructured furnace in Sende Speciiliby Misscoor

Structural input-on vitulity estimation using perovskap-reckson


Drop-deformation and motion analysis of Seltonian are School Documer airk Perfined Worlar keady with planner ziting with distributed to genes and into a virtuad shipblaces

Here are a few that turned out sounding rather sinister. If you were in an scifi novel and an AI was working on these projects, it might be a good time to go hide out on a remote backwater planet.

Ocean disording Scheme

Estimation of embedded orbit radio sequence for human motivation

Projectively controlling systems : the effect of demands-/superxound aircraft engines in uncontent environment

Biological treatment system of compressible virtual transmission

Blockfluid assimilation

Development of the intra-elastic net ecchonic cell bisysties to reality garden’s cell simulator

It’s fun to turn the creativity level higher, making the neural network invent ever more daring thesis titles.

Study of cadmium shall Cju-padence electrodels for social ultrafastering

Measuring cell movements for fluorine ecology

Decisionalization tools through parents for a bacterially and flexure thermor tungsting

Frook.

Spacecraft avoidance : a short

Design for transformating and swirl-thermogravic and efficient congestion

The maneuvership of Bostolk Third electrodelay tissue

Statistical validity of Bell thermophotonic integrated gyn study of civilable and automobile disters and preschools

Nurler.

And it’s also fun to turn the creativity level near zero and find out the title of the quintessential thesis:

A study of the design of a computer aid to structure and the effect of a computer aided selection of the production of the design of a computer aided selection of the production in a stratified control of a stratified flow of the production of the production of the production of the structure of the production of the production of the production of the production of the control of a metallocycle control of a stratified flow of the production of the production of the production of the production of the production of the structure of the presence of a metalloporphous polymer control of a computer aided selection of the structure of the design of a stratified control of a medical device

Become an AI Weirdness supporter to get bonus content! Or become a free subscriber to get new AI Weirdness posts in your inbox.

Subscribe now

Subscribe now