Paused papers
Note: If I were still in academia, this section would be called "Working
papers." As I have left academia, I left these papers in limbo. Please contact me
if you are interested in the last mile effort of taking them to publication.
Non-Determinism in TensorFlow ResNets
ArXiv (2019), with Matthew Willetts
-
We show that the stochasticity in training ResNets for image
classification on GPUs in TensorFlow is dominated by the
non-determinism from GPUs, rather than by the initialisation of
the weights and biases of the network or by the sequence of
minibatches given. The standard deviation of test set accuracy is
0.02 with fixed seeds, compared to 0.027 with different
seeds---nearly 74% of the standard deviation of a ResNet model is
non-deterministic. For test set loss the ratio of standard
deviations is more than 80%. These results call for more robust
evaluation strategies of deep learning models, as a significant
amount of the variation in results across runs can arise simply
from GPU randomness.
The labor market consequences of electricity adoption: concrete evidence from the Great Depression Revise and re-submit to the American Economic Journal: Macroeconomics. (November 2015, INET Working Paper).
Media coverage:
Financial Times,
Royal Economic Society.
- Do labor-saving technologies lead to a fall in employment or an increase in output in the medium-run? This paper estimates the effects of cheaper electricity using the share of coal power as an instrument and the labor market outcomes from the non-traded industry of concrete. Electricity caused electric capital intensity and labor productivity to increase, the labor share of income and employment to decrease, and had no effect on output. The effects are stronger in counties where the Depression was more severe. A structural estimation finds that the elasticity of substitution between electric capital and dexterity tasks at 2.2.
Computer Adoption and the Changing Labor Market Reject and re-submit at the Review of Economic Dynamics. (February 2016).
Media coverage:
Marginal Revolution.
- This paper examines computer adoption as a theoretical explanation for changes in the US labor market in recent decades. When computers become cheap and competitive compared to labor services, they diffuse more rapidly in the conventional mechanism of capital-labor substitution. The model accounts for recent structural changes with this trend of automation: employment shifts away from routine occupations and the labor share of income declines. With hiring costs, firms entering a recession “front-load” the destruction of routine jobs, which accounts for recent cyclical changes of the labor market: routine job losses are concentrated in recessions and the ensuing recoveries are jobless. This paper also tests this labor demand mechanism against the labor supply mechanism of Jaimovich and Siu (2012): computer adoption predicts job layoffs but not job quits among the unemployed.
Failed papers
Tablet language stimulation and sight-reading applications
to narrow toddlers' literacy gap (with Sriya
Iyer). Research proposal. Ethics approval granted.
Note: I gave up on this project for the time being because the early
results from the EmotionAthletes project show a lot more promise.
-
Policymakers and scholars alike worry over the learning gap of children: by age 3, low-income children hear 71% fewer words than high-income children (Hart and Risley, 2003). Early vocabulary knowledge predicts reading ability in later years (Berlinski and Schady, 2015). This gap arises early, is persistent, and has widespread consequences for schooling, criminality, and health (Elango et al., 2015). Scholars have tried to remedy this gap with early childhood non-reading interventions (Attanasio et al., 2015), late childhood reading interventions (Marulis and Neuman, 2010), or early childhood peripheral reading interventions, such as letter awareness or home-based shared reading, (National Early Literacy Panel, 2008). These interventions give moderate to large effect sizes (Elango et al., 2015).
The Effect of Infrastructure Investment in a Low-Growth Environment: Evidence from the Great Depression (with Scott Swisher) Rejected by Explorations in Economic History.
-
We inquire into the effects of infrastructure spending and road construction on the development of the manufacturing sector during the Great Depression. We use a novel dataset on road construction and manufacturing firms between 1930 and 1940. We distinguish between local effects, measured with the change in road mileage by county, and global effects, measured with the change in market access to all other counties through the road network. We also distinguish between the whole manufacturing sector and the least tradable industries. We find that global market access is correlated with an expansion of manufacturing (increase in number of firms, firm size, firm output, and a decrease in average labor productivity) for the whole manufacturing sector and not for the least tradable industries. We also find that the change in local market access has no correlation with manufacturing outcomes, either all of manufacturing or the least tradable industries.
Routinization and Slow Recoveries in
Consumption (3rd
chapter of doctoral
dissertation). Not submitted.
